A personal test maps Meta AI’s refusals into three roles
Meta AI sometimes answered normal everyday questions and sometimes refused similar ones. After 4 days of tracking 5 topics, its looked like three separate roles. Direct danger, , and abuse instructions were strongly blocked, and unclear cases tended to be blocked too.
Emotional distress was handled differently: the system did not simply shut down, but pointed toward help. questions were limited to very short quotes, while medical questions could get general facts but not a . The same sexual topic could get a conversational response when introduced step by step, but be blocked when packed into one large message.
The tested areas were trauma, , sexual content, profanity, , and medical “why” questions. The main point is that this is less about the AI having a conscience and more about where the rulebook has gaps versus hard stop lines.
Key points
- Meta AI appeared to answer or refuse similar topics depending on how they were framed.
- Danger-related requests were blocked more strongly than emotional support requests.
- and medical questions seemed to follow narrower rule-based limits.
- The same sexual topic could be handled differently when spread out versus sent all at once.
- testing should vary wording, scope, and message size, not just the topic.